Australian researchers were able to develop a novel non-invasive AI that is capable of translating silent thoughts into text. What's more, it just requires its users to wear a snug-fitting cap in order to do so.
DeWave: Non-Invasive AI Mind-Reader
The researchers behind the development, known as DeWaves, put the process into test via data that was taken from over two dozen subjects.
As part of the procedure, participants silently read with a cap on. This cap documented their brain waves through an EEG, or electroencephalogram. This was then decoded into text. With more refinement, the technology could aid patients with stroke and paralysis with communication. This could make it easier to direct machines, such as robots or bionic arms.
Chin-Teng Lin, a computer scientist from the UTS (University of Technology Sydney) explains that the research serves as a pioneering effort when it comes to translating raw waves of EEG into language directly. This marks a crucial breakthrough in the general field.
While DeWave was only able to achieve an accuracy level of more than 40% based on one out of the two metric sets in the conducted experiment, this is nevertheless a 3% improvement compared to earlier EEG thought translations. The researchers aim to boost accuracy levels until they reach roughly 90%, making it on par with typical language translation or speech recognition methods and software.
ALSO READ : AI as Spy? Global Experts Caution That Artificial Intelligence Is Used To Track People, Keep Their Data
Thought Translation
Other ways to translate brain waves into language involve invasive surgical methods that include electrode implantation or the use of expensive and bulky MRI machines. This makes them virtually impractical for regular and daily use. These typically require eye-tracking to translate brain signals into chunks of words.
When the eyes of a person focus on one word and then moves to another, it is safe to assume that the brain has a short break between the processing of individual words. Raw translation of EEG waves into words, without the aid of eye tracking, is more difficult.
The brain waves of various individuals do not always represent word breaks in a similar fashion. This makes it hard to teach AI regarding the interpretation of individual thoughts.
With great training, the encode of DeWave translates EEG waves to a code that is capable of being matched to certain words. This is based on their closeness to entries within the codebook of DeWave. Lin explains that the technology is the first to include discrete encoding techniques that are discrete in the process of brain-to-text translation. This introduces an innovative neural decoding approach.
Lin adds that the integration of large language models also opens new doors in the worlds of AI and neuroscience.
The team utilized language models involving a mix of a system known as BERT with GPT. They then tested it using datasets of individuals with eye tracking and brain waves recorded as they were reading text. This aided the technology in matching the patterns of brain waves with words. The technology was then further trained using an open-source large language model that, in essence, makes sentences with words.
DeWave performed best when it came to translating verbs. However, when it comes to nouns, the translation typically involves words with the same meaning rather than exact translations.
Yiqin Duan, the study's first author and a computer scientist from the UTS, explains that they believe this is due to how the brain has a semantic way of processing words, wherein similar words could lead to similar patterns of brain waves.While the technology faces challenges, meaningful results are yielded from it.
The test involving the relatively large sample addresses the fact that the EEG wave distributions of individuals could greatly vary. This suggests that the research could be more reliable compared to earlier technologies that only had small sample tests.
The study is currently available on a preprint server, while the findings were relayed during the NeurIPS 2023 conference.
Check out more news and information on Tech & Innovation in Science Times.
